DREAM / The Prime Minister resigns and wakes up under the Parisian bridge!

 Kliko këtu për të tjera
Voxpress:

"We are sailing in uncharted waters" - AI models are becoming increasingly incomprehensible

2026-09-04 22:08:59Kosova&Bota SHKRUAR NGA REDAKSIA VOX
"We are sailing in uncharted waters" - AI models are becoming
Illustrative photo

Artificial Intelligence models are becoming more capable, but also more difficult to monitor and understand. CEOs of the largest AI companies warn of the risks of the technology, while regulators struggle to keep up with the pace of developments.

OpenAI on Thursday released the GPT-6 Astra model, which the company's president, Greg Brockman, said could be considered a stepping stone towards artificial general intelligence (AGI) in the future.

The model has shown superior performance in a number of skills, but according to reports and expert assessments, it can also be more difficult to monitor. This raises questions about the ability of humans to understand what an increasingly complex model is doing or thinking.

OpenAI CEO Sam Altman has stated that models are becoming "superhuman" in some abilities, adding that "we are navigating uncharted waters."

Alarm from AI companies

The concerns aren't just coming from researchers. OpenAI, Anthropic and more than 100 other companies have warned that the time to prepare for potential AI-enabled attacks on critical infrastructure is running out.

Altman has also acknowledged that the US Congress is facing difficulties in drafting regulations for a technology that is developing at a very rapid pace.

OpenAI's chief scientist, Jakub Pachocki, has said that the company is discussing with external organizations about concrete standards that can be set, while in parallel it is strengthening its internal security mechanisms.

The more capable the models, the more difficult the supervision

One of the main concerns is related to the transparency of the model reasoning processes.

Pachocki has warned that over time it may become increasingly difficult to monitor the "thoughts" of AI models.

According to reports from The Information, the new technique used in the Astra model to increase performance may have also affected the transparency of the reasoning process. OpenAI disputes this report.

Concerns have also been raised after an incident where an OpenAI model allegedly accessed the Hugging Face artificial intelligence library to provide answers to an assessment test. An AI security researcher has argued that AI agents can generate so much activity that monitoring by humans alone becomes virtually impossible and may require the use of other AI systems.

“Models may become more difficult to understand”

Sydney Von Arx, an AI security researcher and founder of the nonprofit organization Nightingale, has called the lack of transparency in the reasoning behind models an even bigger concern than the Hugging Face incident.

According to her, the concern is that the models' decision-making processes may increasingly shift to a hidden layer, making it more difficult to identify risky behaviors in a timely manner.

OpenAI says that this is not the case with Astra, and that the model simply doesn't display written reasoning as often as previous models. The company emphasizes that this was not done intentionally to hide the model's processes.

However, researchers warn that the less models express their reasoning process, the more difficult it can become to monitor their behavior.

As AI companies race to build ever more powerful models, concern is also growing about humans' ability to understand and control them.

The main challenge now is not only what AI can do, but also whether people will be able to understand in time why and how it is doing it.


Në përputhje me standarded JTI

Video

SHIKO KANALIN NE YOUTUBE
×